HOCIntelligentTechnologyGroup
Faculty, BIG DATA (information), Large language model(LLM), Generative Pre-trained Transforme model(GPT), Research Lab's "FBLGR Quality" is a leading company,AI Science and Technology Innovation for Global Sustainable Development.
DX, ICT, AI, Agriculture, Medical, Human Resource Development, Future Prediction, SDGs, etc.
TAHAR
Jun 30, 2023
HOCITGROUP Launch of KOKOGPT, a sentence generation AI service based on "correct data"
HOCITGROUP Launch of KOKOGPT, a sentence generation AI service based on "correct data"
As the reliability of answers becomes important in the use of generative AI, HOCITGROUP has announced that it will launch a service that accurately answers questions based on the data that users want to know.
Development of a large-scale language model (generative AI) specialized for Japanese
~Development of a large-scale language model with 40 billion parameters trained only on Japanese web data~
Using only 350 GB of Japanese Web text collected independently, we developed a large-scale language model of the generation system with 40 billion parameters. Through this development, we have gained a lot of knowledge in the development of large-scale language models of generative systems, such as formatting and filtering of pre-training texts, and pre-training using a large-scale computing infrastructure. Currently, we are training a large-scale language model (equivalent to OpenAI's GPT-3) with 179 billion parameters, and we are also working on the scale of training textbooks. In the future, we plan to collaborate with companies, national research institutes, universities, etc. through joint research and other activities to research and develop and utilize large-scale language models of Japanese.
Generative AI such as "Chat GPT" is rapidly expanding its use around the world.
However, there are challenges where the answers to the questions are wrong or the answers are abstract.
Under these circumstances, HOCITGROUP will start providing a new generation AI service that links data that serves as a reference for answers and provides well-founded answers based on it.
For example, if you ask Chat GPT how to take time off from the company when your child is hospitalized, you will get a general and abstract answer, although it is not wrong, such as "It depends on work rules and working conditions."
On the other hand, if you select the data of the work rules as a reference destination with the generated AI service developed by HOCITGROUP and ask the same question, the user will answer more specifically based on the work rules, such as the procedure method and the impact on salary, after saying that you can take nursing leave.
"Normal chat GPT has the characteristic of telling a plausible lie, which is called halcination, and if you take the form of this document reference type, you will search for internal documents and generate answers based on what is written in the searched documents, so the answers will also be true."
HOCITGROUP provides such generative AI services for enterprises.